Google DeepMind's Co-Scientist reaches Nature publication as a multi-agent research tool

Google DeepMind’s AI co-scientist, first shown to a small group of trusted testers in February 2025, moved from preview to peer-reviewed publication with a paper in Nature describing it as a multi-agent research partner built on Gemini models. The system generates candidate hypotheses, has virtual agents debate and rank them through simulated peer review, and iteratively refines the top candidates before producing a final proposal — a pipeline DeepMind described as Generate, Debate and Evolve.

DeepMind reported six case studies as evidence the tool produced results scientists judged useful rather than merely plausible-sounding: candidate drugs proposed for liver fibrosis were reported to block 91% of a scarring-related cellular response in laboratory tests, and researchers working on cellular ageing said the system cut the time to identify promising genetic leads from months to days. Other case studies covered ALS research, infectious-disease protein targets and metabolic disease.

Across the reported collaborations, DeepMind and the outside researchers involved characterised the system as augmenting scientists’ own hypothesis generation rather than working autonomously, and the published results depended on wet-lab confirmation by human researchers rather than the model’s output alone. The move to a peer-reviewed Nature publication gave the tool a stronger evidentiary basis than the February 2025 preview, though the case studies were selected examples rather than a systematic trial against a control group.